1. Multithrading in CSEARCH

aqme/csearch/base.py

@@ -302,7 +302,7 @@ def __init__(self, **kwargs):
  self.args.log.write(f"\nStarting CSEARCH with {len(job_inputs)} job(s) (SDF, XYZ, CSV, etc. files might contain multiple jobs/structures inside)\n")
 
  # runs the conformer sampling with multiprocessors
- self.run_csearch(job_inputs)
+ _ = self.run_csearch(job_inputs)
 
  # store all the information into a CSV file
  csearch_file_no_path = (
@@ -388,71 +388,48 @@ def run_csearch(self, job_inputs):
  "o Number of finished jobs from CSEARCH", max=len(job_inputs)
  )
 
- # rdkit benefits from using multithreading (the RMSD filter only uses 1 proc, when trying to use
- # more the program collapses)
+ # rdkit benefits from using multithreading, since the RMSD filter in RDKit's GetBestRMS 
+ # doesn't parallelize well (by default, it uses 1 thread and it fails when using more, 
+ # and from our experience this function isn't efficient as we're not sure that
+ # it tries to use all the CPUs or only 1)
  if self.args.program.lower() == "rdkit":
  csearch_procs = self.args.nprocs
- else:
+ else: # CREST already parallelizes CPUs
  csearch_procs = 1
-
- with futures.ProcessPoolExecutor(
+
+ # asynchronous multithreading to accelerate CSEARCH (only benefits RDKit)
+ with futures.ThreadPoolExecutor(
  max_workers=csearch_procs,
  ) as executor:
- # Submit a set of asynchronous jobs
- jobs = []
- # Submit the Jobs
  for job_input in job_inputs:
- (
- smi_,
- name_,
- charge_,
- mult_,
- constraints_atoms_,
- constraints_dist_,
- constraints_angle_,
- constraints_dihedral_,
- complex_type_,
- geom_
- ) = job_input
- job = executor.submit(
- self.compute_confs(
- smi_,
- name_,
- charge_,
- mult_,
- constraints_atoms_,
- constraints_dist_,
- constraints_angle_,
- constraints_dihedral_,
- complex_type_,
- geom_
- )
+ _ = executor.submit(
+ self.compute_confs, job_input,bar
  )
- jobs.append(job)
-
- bar.next()
 
-  bar.finish()
+ bar.finish()
 
- def compute_confs(
- self,
- smi,
- name,
- charge,
- mult,
- constraints_atoms,
- constraints_dist,
- constraints_angle,
- constraints_dihedral,
- complex_type,
- geom
- ):
+ def compute_confs(self,job_input,bar):
  """
  Function to start conformer generation
  """
 
+ # load variables from job_input
+ (
+ smi,
+ name,
+ charge,
+ mult,
+ constraints_atoms,
+ constraints_dist,
+ constraints_angle,
+ constraints_dihedral,
+ complex_type,
+ geom
+ ) = job_input
+
  self.args.log.write(f"\n ----- {os.path.basename(Path(name))} -----")
 
+ # load mol and other parameters when using SMILES as input
  if self.args.smi is not None or os.path.basename(Path(self.args.input)).split(".")[1] in ["smi","csv","cdx","txt","yaml","yml","rtf"]:
  (
  mol,
@@ -477,6 +454,7 @@ def compute_confs(
  if os.path.basename(Path(self.args.input)).split(".")[1] not in ["csv","cdx","txt","yaml","yml","rtf"]:
  self.args.log.finalize()
  sys.exit()
+ bar.next()
  return
 
  else:
@@ -487,6 +465,7 @@ def compute_confs(
  if os.path.basename(Path(self.args.input)).split(".")[1] not in ["csv","cdx","txt","yaml","yml","rtf"]:
  self.args.log.finalize()
  sys.exit()
+ bar.next()
  return
 
  # check if the optimization is constrained
@@ -551,6 +530,7 @@ def compute_confs(
  if os.path.basename(Path(self.args.input)).split(".")[1] not in ["csv","cdx","txt","yaml","yml","rtf"]:
  self.args.log.finalize()
  sys.exit()
+ bar.next()
  return
 
  if complex_type in accepted_complex_types:
@@ -620,6 +600,7 @@ def compute_confs(
  # Updates the dataframe with infromation about conformer generation
  frames = [self.final_dup_data, total_data]
  self.final_dup_data = pd.concat(frames, ignore_index=True, sort=True)
+ bar.next()
 
  # automatic detection of metal atoms 
  def find_metal_atom(self,mol,charge,mult):

aqme/qdescp.py

@@ -374,7 +374,7 @@ def gather_files_and_run(self, destination, atom_props, update_atom_props, smart
  bar = IncrementalBar(
  "\no Number of finished jobs from QDESCP", max=len(self.args.files)
  )
- # multiprocessing to accelerate QDESCP (since xTB uses 1 processor to be reproducible)
+ # asynchronous multithreading to accelerate QDESCP (since xTB uses 1 processor to be reproducible)
  with futures.ThreadPoolExecutor(
  max_workers=self.args.nprocs,
  ) as executor:

aqme/utils.py

@@ -309,7 +309,7 @@ def get_conf_RMS(mol1, mol2, c1, c2, heavy, max_matches_rmsd):
  if heavy:
  mol1 = RemoveHs(mol1)
  mol2 = RemoveHs(mol2)
- return GetBestRMS(mol1, mol2, c1, c2, maxMatches=max_matches_rmsd) # don't use numThreads=0 or -1 as the documentation says, it fails! (due to multiprocessing?)
+ return GetBestRMS(mol1, mol2, c1, c2, maxMatches=max_matches_rmsd) # don't use numThreads=0 or -1 as the documentation says, it fails! (due to multithreading in CSEARCH module?)
 
 
 def command_line_args():